jinzd-ai-cli 0.4.320 → 0.4.322

package/README.md

@@ -44,7 +44,7 @@
 - **Web UI Memory Panel** *(v0.4.90+, B4)* — new 🧠 Memory sidebar tab with semantic search across past chats; each hit has **➕ Inject** (quotes the snippet into the chat input as a markdown blockquote so you can review/edit before sending — no silent context injection) and **↗ Load** (jumps to source session). Bulk "Inject top 3" for recall bundles
 - **Streaming Tool Use** — Real-time streaming of AI reasoning and tool calls as they happen
 - **Sub-Agents** — Delegate complex subtasks to isolated child agents with independent tool loops
-- **Extended Thinking** — Claude deep reasoning mode with `/think` toggle
+- **Thinking control** — `/think on|off|low|high|max` across Claude, DeepSeek, GLM-5, Ollama thinking models (off by default where the API allows it)
 - **Plan Mode** — Read-only planning phase (`/plan`) where AI analyzes before executing, with loop detection
 - **Auto-Pause** — Automatically pauses every 10 rounds for user review and redirection
 - **MCP Protocol** — Connect external MCP servers for dynamic tool discovery
@@ -206,7 +206,7 @@ AI autonomously invokes these 30 tools during conversations:
 | `/provider` | Switch AI provider |
 | `/model` | Switch model; `/model refresh [provider]` refreshes live model cache |
 | `/plan` | Enter read-only planning mode |
-| `/think` | Toggle Claude extended thinking |
+| `/think` | Thinking on/off and effort (`on\|off\|low\|high\|max\|status`) |
 | `/test` | Auto-detect and run project tests |
 | `/review` | AI code review of current git diff |
 | `/security-review` | Security vulnerability scan on git diff |

package/README.zh-CN.md

@@ -34,7 +34,7 @@
 - **Web UI Memory 面板**（v0.4.90+，B4）— sidebar 新增 🧠 Memory 标签页，跨 session 语义搜索；每条 hit 带 **➕ Inject**（把片段作为 markdown 引用块塞进聊天输入框，用户可在上面继续打字编辑——不是静默注入上下文）和 **↗ Load**（跳转对应 session）；顶部"➕ Inject top 3"一键批量
 - **流式工具调用** — 实时流式展示 AI 推理过程和工具调用
 - **子代理系统** — 将复杂子任务委派给独立子代理执行
-- **深度推理** — Claude Extended Thinking，`/think` 一键切换
+- **思考控制** — `/think on|off|low|high|max`，覆盖 Claude、DeepSeek、GLM-5、Ollama 思考模型（API 允许关闭的默认关闭）
 - **规划模式** — `/plan` 进入只读规划，AI 先分析后执行，内置死循环检测
 - **自动暂停** — 每 10 轮自动暂停，用户可审查进度或重定向 AI
 - **MCP 协议** — 接入外部 MCP 服务器，动态发现工具
@@ -203,7 +203,7 @@ AI 在对话中可自主调用 30 个工具：
 | `/provider` | 切换 AI Provider |
 | `/model` | 切换模型 |
 | `/plan` | 进入只读规划模式 |
-| `/think` | 切换 Claude 深度推理 |
+| `/think` | 思考开关与档位（`on\|off\|low\|high\|max\|status`） |
 | `/test` | 自动检测并运行测试 |
 | `/review` | AI 代码审查 git diff |
 | `/security-review` | 安全漏洞扫描 git diff |

package/dist/{batch-LRYDPB6F.js → batch-IMC6Q63Y.js}

@@ -4,12 +4,12 @@ import {
 } from "./chunk-RS7SKYTS.js";
 import {
   ConfigManager
-} from "./chunk-DB4HVUPI.js";
+} from "./chunk-OR46O57D.js";
 import "./chunk-TU3L3PW7.js";
 import {
   atomicWriteFileSync
 } from "./chunk-BMLQPSHP.js";
-import "./chunk-J7IPNING.js";
+import "./chunk-ZUONW7FS.js";
 
 // src/cli/batch.ts
 import Anthropic from "@anthropic-ai/sdk";

package/dist/{chunk-NOAWUEJJ.js → chunk-3TWEKDHH.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   TEST_TIMEOUT
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 import {
   getExecutionConfigDir,
   getExecutionSignal

package/dist/{chunk-TKBWA3XJ.js → chunk-FHBV5EUM.js}

@@ -2,7 +2,7 @@
 import {
   evaluateCiGate,
   normalizeCiThresholds
-} from "./chunk-P4FJBD2H.js";
+} from "./chunk-W7D4M4KV.js";
 import {
   buildReviewPrompt,
   buildSecurityReviewPrompt,
@@ -13,13 +13,13 @@ import {
 } from "./chunk-6YZEZFX5.js";
 import {
   ProviderRegistry
-} from "./chunk-YAMPLTYO.js";
+} from "./chunk-XVB4RARX.js";
 import {
   ConfigManager
-} from "./chunk-DB4HVUPI.js";
+} from "./chunk-OR46O57D.js";
 import {
   VERSION
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/cli/ci.ts
 import { execFileSync } from "child_process";

package/dist/{chunk-WY4IWCQ5.js → chunk-G46THFJN.js}

@@ -13,7 +13,7 @@ import {
   MEMORY_FILE_NAME,
   MEMORY_MAX_CHARS,
   MEMORY_STORE_FILE_NAME
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/memory/persistent-memory.ts
 import { copyFileSync, existsSync, mkdirSync, readFileSync, statSync } from "fs";

package/dist/{chunk-4S2XD3M2.js → chunk-HVVDGSUN.js}

@@ -8,13 +8,13 @@ import {
   listHooks,
   listInstalledPlugins,
   resetStats
-} from "./chunk-5VZHSSLW.js";
+} from "./chunk-RO67VWKW.js";
 import {
   ProviderRegistry
-} from "./chunk-YAMPLTYO.js";
+} from "./chunk-XVB4RARX.js";
 import {
   ConfigManager
-} from "./chunk-DB4HVUPI.js";
+} from "./chunk-OR46O57D.js";
 import {
   getGitRoot
 } from "./chunk-HOSJZMQS.js";
@@ -26,7 +26,7 @@ import {
   MCP_PROJECT_CONFIG_NAME,
   MEMORY_FILE_NAME,
   VERSION
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/diagnostics/doctor-report.ts
 import { existsSync as existsSync3, statSync as statSync2 } from "fs";

package/dist/{chunk-46OEMKQX.js → chunk-NFPJ6KUU.js}

@@ -108,17 +108,21 @@ var PRICING_TABLE = {
   "glm-4.6": { input: 0.6, output: 2.2 },
   "glm-4.6v": { input: 0.6, output: 2.2 },
   "glm-5": { input: 0.85, output: 2.85 },
-  // GLM-5.3（2026-08-14 发布）官方公告口径「与 GLM-5.2 同价」，直连 API $1.40 / $4.40。
+  // GLM-5.3：官方定价页（open.bigmodel.cn/pricing，2026-10-08 浏览器读取）——输入 8 元、
+  // 输出 28 元、缓存命中 2 元（每百万 token），按 ¥6.9/$1 换算。此前记的 $1.40/$4.40 是
+  // Z.ai 国际站的美元价，而本 provider 打的是国内站、按人民币计费，高报约 20%；更大的
+  // 偏差在**缓存**：没有 cacheRead 时缓存命中按输入全价算，2026-10-07 审计会话 1474 万
+  // 缓存 token 被记成 $21.34，按官方缓存价约 $4.90（2026-10-07 审计 X-03）。
   // 注意它**始终思考**，思考 token 按输出价计费 —— 实际成本高度取决于 reasoning_effort。
-  "glm-5.3": { input: 1.4, output: 4.4 },
+  "glm-5.3": { input: 8 / CNY_PER_USD, output: 28 / CNY_PER_USD, cacheRead: 2 / CNY_PER_USD },
   // GLM-5.3-Flash：官方定价页（open.bigmodel.cn/pricing）一手数字——输入 0.8 元、
   // 输出 2.8 元、缓存命中 0.23 元（每百万 token），按记账汇率 ¥6.9/$1 换算。
   // **刻意记恢复价而非当前的 5 折限时价**（0.4/1.4/0.115 元）：官方只说「限时两周」、
   // 没公布截止日期，静态表表达不了这个时间维度；高估比低估安全——预算告警早触发
   // 好过晚触发。折扣期内 `/cost` 会比实际偏高约一倍，这是刻意的取舍。
   "glm-5.3-flash": { input: 0.116, output: 0.406, cacheRead: 0.033 },
-  // GLM-5.2（2026-06）官方直连 API 定价 $1.40 / $4.40。
-  "glm-5.2": { input: 1.4, output: 4.4 },
+  // GLM-5.2：与 GLM-5.3 同价（同一定价页：8 / 28 / 缓存命中 2 元），口径修正同上。
+  "glm-5.2": { input: 8 / CNY_PER_USD, output: 28 / CNY_PER_USD, cacheRead: 2 / CNY_PER_USD },
   // GLM-5V-Turbo（2026-04 视觉编程旗舰）官方直连 API 定价 $1.20 / $4.00。
   "glm-5v-turbo": { input: 1.2, output: 4 },
   "glm-5.1": { input: 0.95, output: 3.15 },

package/dist/{chunk-WW2LFOWE.js → chunk-NX72SBGU.js}

@@ -12,7 +12,7 @@ import {
   MEMORY_FILE_NAME,
   MEMORY_MAX_CHARS,
   MEMORY_STORE_FILE_NAME
-} from "./chunk-SKQUWDRM.js";
+} from "./chunk-X5ZC4TI4.js";
 
 // src/memory/persistent-memory.ts
 import { copyFileSync, existsSync, mkdirSync, readFileSync, statSync } from "fs";

package/dist/{chunk-DB4HVUPI.js → chunk-OR46O57D.js}

@@ -11,7 +11,7 @@ import {
   CONFIG_FILE_NAME,
   HISTORY_DIR_NAME,
   PLUGINS_DIR_NAME
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/config/config-manager.ts
 import { readFileSync, existsSync, mkdirSync } from "fs";

package/dist/{chunk-5VZHSSLW.js → chunk-RO67VWKW.js}

@@ -8,7 +8,7 @@ import {
 import {
   CONFIG_DIR_NAME,
   PLUGINS_DIR_NAME
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/diagnostics/tool-stats.ts
 import { mkdirSync } from "fs";

package/dist/{chunk-N6YFBRQV.js → chunk-RSWCASRV.js}

@@ -4,7 +4,7 @@ import {
 } from "./chunk-QMFMST2T.js";
 import {
   TEST_TIMEOUT
-} from "./chunk-SKQUWDRM.js";
+} from "./chunk-X5ZC4TI4.js";
 
 // src/tools/builtin/run-tests.ts
 import { spawn as spawn2 } from "child_process";

package/dist/{chunk-QB3MECZJ.js → chunk-SYK2TDKV.js}

@@ -2,7 +2,7 @@
 import {
   computeCost,
   formatCost
-} from "./chunk-46OEMKQX.js";
+} from "./chunk-NFPJ6KUU.js";
 import {
   clearRecentlyDeniedAutoActions,
   estimateTokens,
@@ -16,7 +16,7 @@ import {
   setPreferredAgentName,
   truncateForPersist,
   undoStack
-} from "./chunk-46M65JDN.js";
+} from "./chunk-ZSR2UIRD.js";
 import {
   buildReviewPrompt,
   buildSecurityReviewPrompt
@@ -51,7 +51,7 @@ import {
   touchMemoryReferences,
   updateMemoryApproval,
   updateMemoryEntry
-} from "./chunk-WY4IWCQ5.js";
+} from "./chunk-G46THFJN.js";
 import {
   theme
 } from "./chunk-5C76XDDF.js";
@@ -65,7 +65,7 @@ import {
 import {
   buildDoctorReport,
   formatDoctorReport
-} from "./chunk-4S2XD3M2.js";
+} from "./chunk-HVVDGSUN.js";
 import {
   describePlugin,
   getActivePluginAssets,
@@ -80,7 +80,7 @@ import {
   trustHook,
   trustPlugin,
   untrustHook
-} from "./chunk-5VZHSSLW.js";
+} from "./chunk-RO67VWKW.js";
 import {
   isCleanDocumentBody,
   stripOuterCodeFence
@@ -109,11 +109,11 @@ import {
   MCP_TOOL_PREFIX,
   REPO_URL,
   VERSION
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/skills/manager.ts
 import { existsSync, readdirSync, mkdirSync, statSync } from "fs";
-import { join } from "path";
+import { basename as basename2, dirname, join } from "path";
 
 // src/skills/types.ts
 import { readFileSync } from "fs";
@@ -253,7 +253,14 @@ var SkillManager = class {
   }
   /** 获取激活技能的 prompt 内容（null 表示无激活技能） */
   getActivePromptContent() {
-    return this.activeSkill?.content ?? null;
+    const skill = this.activeSkill;
+    if (!skill) return null;
+    if (basename2(skill.filePath) === "SKILL.md") {
+      return `Skill directory (SKILL_DIR, \u6280\u80FD\u76EE\u5F55): ${dirname(skill.filePath)}
+
+${skill.content}`;
+    }
+    return skill.content;
   }
   /** 获取激活技能的工具白名单 Set（null 表示无限制） */
   getActiveToolFilter() {
@@ -2426,7 +2433,7 @@ No tools match "${filter}".
           const { join: join6 } = await import("path");
           const { existsSync: existsSync6 } = await import("fs");
           const { getGitRoot } = await import("./git-context-EXOEHQSF.js");
-          const { MCP_PROJECT_CONFIG_NAME } = await import("./constants-P33UDQVS.js");
+          const { MCP_PROJECT_CONFIG_NAME } = await import("./constants-SSVJCBAT.js");
           const { approveProject, hashMcpFile } = await import("./project-trust-KGG2BWKC.js");
           const cwd = process.cwd();
           const projectRoot = getGitRoot(cwd) ?? cwd;
@@ -3534,7 +3541,7 @@ function createReviewCommands() {
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-SYJG4S3N.js");
+          const { executeTests } = await import("./run-tests-WJJW77HK.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {
@@ -4391,7 +4398,7 @@ ${view.chatHeading}
 
 // src/repl/commands/session-commands.ts
 import { writeFileSync as writeFileSync2, mkdirSync as mkdirSync2 } from "fs";
-import { resolve, dirname } from "path";
+import { resolve, dirname as dirname2 } from "path";
 import chalk2 from "chalk";
 
 // src/core/session-cost.ts
@@ -4718,7 +4725,7 @@ function createSessionCommands() {
         const exported = buildSessionExport(session, format, outFile);
         const absPath = resolve(process.cwd(), exported.filename);
         try {
-          mkdirSync2(dirname(absPath), { recursive: true });
+          mkdirSync2(dirname2(absPath), { recursive: true });
           writeFileSync2(absPath, exported.content, "utf-8");
           ctx.renderer.printSuccess(`Exported ${exported.messageCount} messages \u2192 ${absPath}`);
         } catch (err) {
@@ -4885,7 +4892,7 @@ function createSessionCommands() {
 }
 
 // src/repl/commands/misc-commands.ts
-import { basename as basename2 } from "path";
+import { basename as basename3 } from "path";
 import { platform as platform3 } from "os";
 function createMiscCommands() {
   return [
@@ -5142,7 +5149,7 @@ ${hint}` : "")
           return;
         }
         const description = args.join(" ").trim() || "Please describe this image";
-        ctx.renderer.printSuccess(`\u{1F4CB} Image loaded: ${basename2(imgPath)}`);
+        ctx.renderer.printSuccess(`\u{1F4CB} Image loaded: ${basename3(imgPath)}`);
         await ctx.sendAsChat(`@${imgPath} ${description}`);
       }
     },
@@ -5587,7 +5594,7 @@ function clearDevState() {
 
 // src/tools/deferred-save.ts
 import { statSync as statSync4 } from "fs";
-import { basename as basename3 } from "path";
+import { basename as basename4 } from "path";
 function snapshotFile(path) {
   try {
     const st = statSync4(path);
@@ -5608,7 +5615,7 @@ var SAVE_CLAIM_PATTERNS = [
   /\b(?:saved|written|created|generated)\s+(?:to|at|as|into)\b/i
 ];
 function narratesSavedTarget(content, path) {
-  const base = basename3(path);
+  const base = basename4(path);
   if (!base || !content.includes(base)) return null;
   for (const re of SAVE_CLAIM_PATTERNS) {
     if (re.test(content)) return re.source;
@@ -5630,7 +5637,7 @@ function decideDeferredSave(params) {
     return {
       kind: "skip",
       reason: "narrates-target",
-      message: `\u21AA \u8DF3\u8FC7 deferred \u843D\u76D8\uFF1A\u6700\u7EC8\u56DE\u590D\u662F\u5173\u4E8E ${basename3(path)} \u7684\u6C47\u62A5\u800C\u975E\u6587\u6863\u6B63\u6587\uFF08matched: ${narrated}\uFF09\u3002`
+      message: `\u21AA \u8DF3\u8FC7 deferred \u843D\u76D8\uFF1A\u6700\u7EC8\u56DE\u590D\u662F\u5173\u4E8E ${basename4(path)} \u7684\u6C47\u62A5\u800C\u975E\u6587\u6863\u6B63\u6587\uFF08matched: ${narrated}\uFF09\u3002`
     };
   }
   if (!isCleanDocumentBody(body)) {

package/dist/{chunk-P4FJBD2H.js → chunk-W7D4M4KV.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   VERSION
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 
 // src/cli/ci-format.ts
 var DEFAULT_CI_THRESHOLDS = {

package/dist/{chunk-SKQUWDRM.js → chunk-X5ZC4TI4.js}

@@ -1,5 +1,5 @@
 // src/core/constants.ts
-var VERSION = "0.4.320";
+var VERSION = "0.4.322";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-YAMPLTYO.js → chunk-XVB4RARX.js}

@@ -2120,6 +2120,7 @@ function ollamaNativeBaseUrl(openAiBaseUrl) {
   return openAiBaseUrl.replace(/\/+$/, "").replace(/\/v1$/, "");
 }
 var NATIVE_API_TIMEOUT_MS = 3e3;
+var SMALL_CONTEXT_WARN = 16384;
 var OllamaProvider = class extends OpenAICompatibleProvider {
   // 用 127.0.0.1 而非 localhost：Node.js 18+ 在 Windows 上优先解析 localhost 为 IPv6 (::1)，
   // 但 Ollama 默认只监听 IPv4，导致连接失败。
@@ -2198,12 +2199,13 @@ var OllamaProvider = class extends OpenAICompatibleProvider {
       if (show?.capabilities?.includes("thinking")) thinkingModels.add(id);
       if (running.size === 0 && !show) return;
       const digest = digests.get(id);
-      const { contextWindow } = resolveOllamaContextWindow({
+      const { contextWindow, source } = resolveOllamaContextWindow({
         runningContext: digest && running.get(digest) || running.get(id),
         show,
         fallback: this.estimateContextWindow(id)
       });
       result.set(id, contextWindow);
+      this.contextSources.set(id, source);
     }));
     return { contextWindows: result, thinkingModels };
   }
@@ -2299,6 +2301,77 @@ var OllamaProvider = class extends OpenAICompatibleProvider {
     const target = norm(model);
     return this.dynamicModels.some((m) => m.supportsThinking === true && norm(m.id) === target);
   }
+  // ── 加载后补读真实窗口（2026-10-07 审计 C-01）────────────────────────────
+  //
+  // 上下文窗口只在启动时读一次。模型当时若还没加载，`/api/ps` 里没有它，只能取训练上限
+  // （Qwen3.8 是 262K）；而服务端实际开的窗口可能只有它的几十分之一。第一次请求把模型
+  // 加载进来之后若不再读，上下文压力监控、自动压缩与工具预算就整场按那个大数在算——
+  // aicli 以为还有余量，Ollama 早已在悄悄截掉最早的消息（v0.4.311 那次说的正是这个后果）。
+  // 所以每个模型在首次请求之后补读一次 `/api/ps`；读到的窗口若小得连工具定义都放不下，
+  // 提示一次怎么调大。都是后台做，不拖慢这次请求。
+  /** 每个模型的窗口是从哪一级读到的（只有 'running' 是服务端实际值） */
+  contextSources = /* @__PURE__ */ new Map();
+  refreshedAfterLoad = /* @__PURE__ */ new Set();
+  warnedSmallContext = /* @__PURE__ */ new Set();
+  async chat(request) {
+    try {
+      return await super.chat(request);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async *chatStream(request) {
+    try {
+      yield* super.chatStream(request);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async chatWithTools(request, tools) {
+    try {
+      return await super.chatWithTools(request, tools);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async *chatWithToolsStream(request, tools) {
+    try {
+      yield* super.chatWithToolsStream(request, tools);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  scheduleContextRefresh(model) {
+    const key = model.replace(/:latest$/, "");
+    if (this.refreshedAfterLoad.has(key)) return;
+    this.refreshedAfterLoad.add(key);
+    void this.refreshLoadedContext(model).catch(() => {
+    });
+  }
+  /** 补读已加载模型的实际窗口，更新同一份权重（digest 相同）的所有别名。测试可直接 await。 */
+  async refreshLoadedContext(model) {
+    const norm = (id2) => id2.replace(/:latest$/, "");
+    const [running, digests] = await Promise.all([this.fetchRunningContexts(), this.fetchDigests()]);
+    const self = this.dynamicModels.find((m) => norm(m.id) === norm(model));
+    const id = self?.id ?? model;
+    const digest = digests.get(id);
+    const actual = digest && running.get(digest) || running.get(id);
+    if (!actual) return;
+    for (const m of this.dynamicModels) {
+      if (m.id === id || digest && digests.get(m.id) === digest) {
+        m.contextWindow = actual;
+        this.contextSources.set(m.id, "running");
+      }
+    }
+    if (actual < SMALL_CONTEXT_WARN && !this.warnedSmallContext.has(norm(id))) {
+      this.warnedSmallContext.add(norm(id));
+      process.stderr.write(
+        `
+[ollama] ${id} is loaded with a ${actual.toLocaleString()}-token context window. aicli's tool definitions alone take about 8K tokens, so earlier messages will be silently cut off. Raise it: set OLLAMA_CONTEXT_LENGTH=32768 (or more) for the Ollama server, or add "PARAMETER num_ctx 32768" to the model's Modelfile.
+`
+      );
+    }
+  }
   /** 实时刷新模型列表 */
   async listModels() {
     await this.tryConnect();

package/dist/{chunk-46M65JDN.js → chunk-ZSR2UIRD.js}

@@ -11,7 +11,7 @@ import {
   searchPersistentMemories,
   touchMemoryReferences,
   updateMemoryEntry
-} from "./chunk-WY4IWCQ5.js";
+} from "./chunk-G46THFJN.js";
 import {
   indexProject
 } from "./chunk-RWRZVQCB.js";
@@ -30,7 +30,7 @@ import {
   runHook,
   runLifecycleHooks,
   runTool
-} from "./chunk-5VZHSSLW.js";
+} from "./chunk-RO67VWKW.js";
 import {
   buildTruncatedArgumentsError,
   getDangerLevel,
@@ -64,7 +64,7 @@ import {
   recoverGbkMojibake,
   runTestsTool,
   shellInvocation
-} from "./chunk-NOAWUEJJ.js";
+} from "./chunk-3TWEKDHH.js";
 import {
   CONFIG_DIR_NAME,
   DEFAULT_MAX_TOOL_OUTPUT_CHARS_CAP,
@@ -73,7 +73,7 @@ import {
   SUBAGENT_ALLOWED_TOOLS,
   SUBAGENT_DEFAULT_MAX_ROUNDS,
   SUBAGENT_MAX_ROUNDS_LIMIT
-} from "./chunk-J7IPNING.js";
+} from "./chunk-ZUONW7FS.js";
 import {
   getCurrentSessionKey,
   getExecutionConfigDir,
@@ -6872,21 +6872,37 @@ function estimateTokens(text) {
   return Math.ceil(tokens) + 4;
 }
 function estimateToolDefinitionTokens(def) {
-  let charCount = def.name.length + def.description.length;
-  for (const [key, param] of Object.entries(def.parameters)) {
-    charCount += key.length + param.type.length + param.description.length;
-    if (param.enum) {
-      charCount += param.enum.join(" ").length;
-    }
-  }
-  const structuralOverhead = 40 + Object.keys(def.parameters).length * 30;
-  return Math.ceil((charCount + structuralOverhead) * 0.25) + 10;
+  const json = JSON.stringify({ name: def.name, description: def.description, parameters: def.parameters });
+  const cjk = json.match(CJK_REGEX)?.length ?? 0;
+  return Math.ceil((json.length - cjk + 40) / 3.7 + cjk * 1.5);
 }
 
 // src/tools/registry.ts
 import { pathToFileURL } from "url";
 import { existsSync as existsSync13, mkdirSync as mkdirSync3, readdirSync as readdirSync7 } from "fs";
 import { join as join6 } from "path";
+var TRIMMABLE_BUILTIN_ORDER = [
+  "google_search",
+  "recall_memory",
+  "search_code",
+  "notebook_edit",
+  "spawn_agent",
+  "run_interactive",
+  "save_memory",
+  "find_references",
+  "get_outline",
+  "find_symbol",
+  "task_stop",
+  "task_list",
+  "task_create",
+  "git_log",
+  "git_diff",
+  "git_commit",
+  "git_status",
+  "web_fetch",
+  "web_search",
+  "run_tests"
+];
 var ToolRegistry = class {
   tools = /* @__PURE__ */ new Map();
   pluginToolNames = /* @__PURE__ */ new Set();
@@ -6954,7 +6970,7 @@ var ToolRegistry = class {
       totalTokens += estimateToolDefinitionTokens(def);
     }
     if (totalTokens <= tokenBudget) {
-      return { definitions: allDefs, trimmedCount: 0, systemNote: null };
+      return { definitions: allDefs, trimmedCount: 0, trimmedBuiltins: [], systemNote: null };
     }
     const builtinDefs = [];
     const mcpDefs = [];
@@ -6966,10 +6982,19 @@ var ToolRegistry = class {
       }
     }
     let budget = tokenBudget;
-    for (const def of builtinDefs) {
-      budget -= estimateToolDefinitionTokens(def);
-    }
-    const kept = [...builtinDefs];
+    const trimmedBuiltins = [];
+    let builtinCost = builtinDefs.reduce((n, d) => n + estimateToolDefinitionTokens(d), 0);
+    const keptBuiltins = new Set(builtinDefs.map((d) => d.name));
+    for (const name of TRIMMABLE_BUILTIN_ORDER) {
+      if (builtinCost <= budget) break;
+      const def = builtinDefs.find((d) => d.name === name);
+      if (!def || this.pluginToolNames.has(name)) continue;
+      keptBuiltins.delete(name);
+      trimmedBuiltins.push(name);
+      builtinCost -= estimateToolDefinitionTokens(def);
+    }
+    budget -= builtinCost;
+    const kept = builtinDefs.filter((d) => keptBuiltins.has(d.name));
     const used = usedToolNames ?? /* @__PURE__ */ new Set();
     const remaining = [];
     for (const def of mcpDefs) {
@@ -6991,6 +7016,12 @@ var ToolRegistry = class {
         trimmed++;
       }
     }
+    const notes = [];
+    if (trimmedBuiltins.length > 0) {
+      notes.push(
+        `[Tool Budget] To fit this model's context window, these tools are not available in this session: ${trimmedBuiltins.join(", ")}. Use bash for the equivalent (git commands, grep for symbols, curl for web pages) instead of calling them.`
+      );
+    }
     let systemNote = null;
     if (trimmed > 0) {
       const trimmedServers = /* @__PURE__ */ new Set();
@@ -7003,7 +7034,13 @@ var ToolRegistry = class {
       }
       systemNote = `[MCP Tool Budget] ${trimmed} MCP tool(s) from server(s) [${[...trimmedServers].join(", ")}] were excluded to fit the context window. If you need a specific MCP tool that isn't listed, tell the user to ask for it explicitly.`;
     }
-    return { definitions: kept, trimmedCount: trimmed, systemNote };
+    if (systemNote) notes.push(systemNote);
+    return {
+      definitions: kept,
+      trimmedCount: trimmed,
+      trimmedBuiltins,
+      systemNote: notes.length > 0 ? notes.join("\n\n") : null
+    };
   }
   listAll() {
     return [...this.tools.values()];

package/dist/{chunk-J7IPNING.js → chunk-ZUONW7FS.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/core/constants.ts
-var VERSION = "0.4.320";
+var VERSION = "0.4.322";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{ci-7TSOMDHV.js → ci-74EPK5JN.js}

@@ -3,16 +3,16 @@ import {
   CI_COMMENT_MARKER,
   countSeverity,
   runCi
-} from "./chunk-TKBWA3XJ.js";
-import "./chunk-P4FJBD2H.js";
+} from "./chunk-FHBV5EUM.js";
+import "./chunk-W7D4M4KV.js";
 import "./chunk-HLWUDRBO.js";
 import "./chunk-6YZEZFX5.js";
-import "./chunk-YAMPLTYO.js";
-import "./chunk-DB4HVUPI.js";
+import "./chunk-XVB4RARX.js";
+import "./chunk-OR46O57D.js";
 import "./chunk-E2W2MI6H.js";
 import "./chunk-TU3L3PW7.js";
 import "./chunk-BMLQPSHP.js";
-import "./chunk-J7IPNING.js";
+import "./chunk-ZUONW7FS.js";
 export {
   CI_COMMENT_MARKER,
   countSeverity,

package/dist/{ci-format-6A2RFYZS.js → ci-format-QADWGCX7.js}

@@ -6,8 +6,8 @@ import {
   formatCiResult,
   formatCiSarif,
   normalizeCiThresholds
-} from "./chunk-P4FJBD2H.js";
-import "./chunk-J7IPNING.js";
+} from "./chunk-W7D4M4KV.js";
+import "./chunk-ZUONW7FS.js";
 export {
   DEFAULT_CI_THRESHOLDS,
   detectCiGateSignals,

package/dist/{constants-EHYM2NS5.js → constants-P5JJ6FO4.js}

@@ -38,7 +38,7 @@ import {
   TEST_TIMEOUT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-SKQUWDRM.js";
+} from "./chunk-X5ZC4TI4.js";
 export {
   AGENTIC_BEHAVIOR_GUIDELINE,
   APP_NAME,